The Effect of Civilian Casualties on Wartime Informing:Evidence from the Iraq War∗

Andrew Shaver†and Jacob N. Shapiro‡

This draft January 3, 2016

Word Count: 8,812

Abstract

Scholars of civil war and insurgency have long posited that insurgent organizationsand their state enemies incur costs for the collateral damage they cause. We providethe first direct quantitative evidence that wartime informing is affected by civiliancasualties. Using newly declassified data on tip flow to Coalition forces in Iraq we findthat information flow goes down after government forces inadvertently kill civiliansand it goes up when insurgents do so. These results have strong policy implications;confirm a relationship long posited in the theoretical literature on insurgency; and areconsistent with a broad range of circumstantial evidence on the topic.

∗We thank David Carter, Ethan Bueno de Mesquita, Eli Berman, Joe Felter, Frank Gunter, SolomonHsiang, and Christoph Mikulaschek for helpful comments. Participants in the Stanford University’s Inter-national Relations/CISAC Fellows’ Policy Workshop and the National Bureau of Economic Research’s 2015Summer Institute Conference on the Economics of National Security also provided valuable feedback. We areespecially grateful to Jamie Hansen-Lewis and Landin Smith for their contributions to this project. We alsoacknowledge support from AFOSR grant #FA9550-09-1-0314. Numerous officials were extremely generouswith their time in helping us secure the authorized release of the tips data and providing rich context forunderstanding how the tip lines worked in Iraq. All errors are our own. Data used for this article will even-tually be made available at scholar.princeton.edu/ashaver and on the Empirical Studies of Conflict (ESOC)project website.†Woodrow Wilson School, Princeton University, Princeton, NJ 08540, Email: ashaver@princeton.edu‡Woodrow Wilson School, Department of Politics, Princeton University, Princeton, NJ 08540. Email:

jns@princeton.edu. Corresponding author.

1

1 Introduction

Civilians often bear significant costs of fighting between parties to sub-state conflicts. In

Afghanistan, an estimated 21,000 civilians have “died violent deaths as a result of the [on-

going] war” (Costs of War Project, 2014). In Iraq, between 134,789 and 152,104 civilians

have been killed since the United States invasion in 2003 (Iraq Body Count, 2008). The toll

in Syria has been even greater. Price et al. (2014) calculate that nearly 200,000 individuals

have been killed during that country’s four-year old civil war.1 And these numbers do not

include the grievous injuries suffered by many of the conflicts’ civilian survivors. While some

3,200 civilians are reported to have been killed in fighting between the Taliban and Afghan

forces in 2014, more than twice as many Afghans were wounded during the same period

(Johnson, 2014).

Scholars have long posited that insurgent organizations and their state enemies incur costs

for the collateral damage they cause. Most twentieth-century counterinsurgency theorists

writing on the wars of decolonization argued that obtaining information on rebels from

non-combatants was critical for government forces and that protecting the population from

insurgents was critical to gaining that cooperation (Trinquier 1961; Galula 1964; Taber 1965;

Clutterbuck 1966; Thompson 1966; Kitson 1977). More recently, Kalyvas (2006) argued that

indiscriminate violence against civilians is counterproductive because it can turn civilians

against the party causing them harm.2 Berman et al. (2011) explicitly model the relationship

between harm and informing as part of a 3-sided game in which civilians punish insurgents

who create excessive costs by sharing information with government forces.

Indirect empirical support for this relationship has been identified in a number of conflicts.

U.S. government officials cited the potential for civilian casualties to harm cooperation from

1Though this estimate does not distinguish between civilians and combatants, the Syrian Observatoryfor Human Rights contends that 62,347 of these individuals were civilians (Alfred, 2014).

2Though see Lyall (2009) for evidence to the contrary from the Chechen war.

2

civilians as one reason for imposing more restrictive rules of engagement in Afghanistan

in 2009. Condra and Shapiro (2012) show that in Iraq insurgent violence went up after

Coalition-caused civilian casualties and down after insurgent-caused ones, consistent with an

informational reaction to abuse. Lyall et al. (2013a) show that self-reported victimization by

the International Security Assistance Force (ISAF) correlates with lower support for ISAF

and higher support for the Taliban. Shapiro and Weidmann (2015) provide evidence that

plausibly exogenous increases in cell phone coverages led to lower insurgent violence during

the Iraq war, which they attribute to greater information flow to counterinsurgents as the

presence of mobile telecommunications lowered the risks of informing.

What is missing in these papers is direct evidence on information flow to counterinsur-

gents. As one recent paper puts it, “despite its central role in civil war dynamics, the act of

informing is still poorly understood, due mostly to the classified nature of informant tips”

(Lyall et al., 2013b).

Using newly declassified data on weekly province-level tips collected by Iraqi and Coali-

tion force during the Iraq war, we provide the first direct test of the influence of civilian

abuse on wartime informing. Our data span all thirteen provinces which experienced sub-

stantial violence over a 60-week period from June 2007 to July 2008. We combine the data

on tips with administratively-collected geolocated data on combat violence and press-based

data on civilian casualties. We buttress our quantitative analysis with anecdotes from rich

qualitative documentation on intelligence collection in Iraq obtained through Freedom of

Information Act (FOIA) requests.

Exploiting plausibly exogenous variation in civilian casualties occurring during combat

incidents we find a robust relationship between indiscriminate violence and informing. In our

baseline model, an additional Coalition-caused civilian casualty leads to approximately .8

fewer tips in the next week, while an additional insurgent-caused one leads to approximately

.5 more tips. Consistent with prior work we find that government forces pay a higher cost

3

for causing casualties in so far as the drop in tip flows following a single government-caused

casualties is roughly 60% larger than the increase following insurgent caused ones.3 These

effects, while modest in magnitude, are substantively significant. In the median week in which

insurgents caused civilian casualties, they killed four civilians, predicting two additional tips

to Coalition forces. That is a substantial number (roughly 10% of the weekly mean) since

single tips often resulted in raids that led to the capture of both large numbers of weapons

and prominent insurgents.

From a scientific standpoint, these results provide the first direct statistical evidence

for a relationship scholars and practitioners have posited for more than 50 years. One can

interpret our estimates as reflecting a causal effect to the extent that he or she believes

the spikes in civilian casualties from week-to-week are conditionally independent of the next

week’s trend in informing. As we will see, the results below are extremely stable across

different specifications, lending credence to that interpretation.

From a policy perspective, our findings reinforce the importance of minimizing collateral

damage. In addition to the ethical imperative that combatants take all reasonable measures

to avoid harming civilians, they appear to face strategic incentives to do so as well: at least

in Iraq members of the public penalized parties who did not.

This paper proceeds as follows: in the following section, we consider the possible effect(s)

of civilian casualties on battlefield outcomes and why incidents of collateral damage should

be expected to affect wartime informing. We then introduce our data and empirical strategy.

Finally, we present results and conclude with a discussion of policy implications.

3The difference is modestly significant statistically, t=1.48, p=.16, using standard clustered s.e..

4

2 Theory

A growing body of literature analyzes the effects of civilian victimization on battlefield

outcomes during periods of insurgency. The effect of civilian casualties on wartime informing

is of particular interest as scholars of sub-state violence have long highlighted the importance

of information in insurgency campaigns. While we are agnostic as to which mechanism

drives the observed relationship, this section highlights that there are a host of reasons to

expect informing to drop after government forces cause civilians harm and to increase after

insurgents do so.

2.1 The Importance of Informing

A broad range of work argues that incidents of civilian casualties should affect citizens’

willingness to inform. Condra and Shapiro (2012), for example, argue that: “collateral

damage causes local noncombatants to effectively punish the armed group responsible by

sharing more (less) information about insurgents with government forces and their allies

when insurgent (government) forces kill civilians.” This hypothesis emerges from a tradition

of scholarship that focuses on the central role played by civilians in civil war and insurgency

in deciding whether or not to share information about combatants (Tse-Tung (1961), Berman

et al. (2011), Lyall et al. (2013b), Irish Republican Army (1985), Kalyvas et al. (2006), Nagl

(2009)). This line of reasoning emphasizes how the asymmetry of combat power in many

non-state wars renders information about insurgent activities “a central resource in civil

wars: counter-insurgents seek it, insurgents safeguard it, and civilians often trade it” (Lyall

et al., 2013b).

The investment by Iraqi and Coalition forces made in creating anonymous tips channels

provides prima facie evidence that counterinsurgents find tips useful. Telephone tip hotlines

5

like the one shown in figure 1 were active throughout the war.4 The United States established

other platforms as well so that information could be discreetly transmitted to state forces;

these included e-mail addresses as well as online internet web forms like that run by the

Central Intelligence Agency throughout the Iraq war shown in figure 2.

[Insert figures 1 and 2 about here.]

Beyond the logical inference that these systems must have been useful given that they

were maintained for so long, there is substantial evidence that tips served an important

counterinsurgency function in Iraq. Internal government documents released to the authors

describe anonymous tips collection programs during the recent Iraq war as “a critical in-

formation resource for both [Coalition Forces] and [the Government of Iraq] in combating

terrorism” (U.S. Central Command, U.S. Department of Defense, 2007). Field reports link

tips to battlefield success in a number of ways, including leading to: major cache discov-

eries such as “more than 450 deadly anti-tank mines... in the stronghold of support for

radical cleric Muqtada al-Sadr’s militia” (Garamone, 2007); discovery of bomb-making fa-

cilities (Miles, 2007); and arrests of “hundreds of key former Ba’athists and Al Qaeda and

the eventual bombings that finally killed Zarqawi” (Castro, 2007); to highlight just a few.

A national tips hotline jointly run by the U.S. and Iraqi governments “was so effective,

that units began to setup local regional tip centers... [and the] President (US), Vice-President

(US), and SECDEF have all requested historical data for the program” (Multi-National

Corps – Iraq, 2008). While it is unclear what proportion of operations carried out were based

on tips received, those organizations that reported integrating tips into their intelligence cy-

cle claimed they relied heavily on the channel. The tips program reportedly “provide[d]

[Multi National Forces-Iraq’s Strategic Counterintelligence Directorate (SCID)] with 80% of

its [human intelligence] sourcing and 100% of [its] operations” (U.S. Central Command, U.S.

4Indeed, one prominent counterinsurgent predicted in the 1960s governments would find it worthwhile toprovide “would-be informers... a safe, anonymous way to convey information” (Galula, 2006).

6

Department of Defense, 2007), and was considered “very successful” by American special op-

erations forces targeting high priority individuals (U.S. Central Command, U.S. Department

of Defense, 2007).

More recently, following the Islamic State’s advances into Iraq, the country’s central

government reestablished tips lines so that residents of “Fallujah and Ramadi” may “report...

suspected terrorist operations” (Al Shorfa, 2014, 2013; al Qaisi, 2014). In Afghanistan, less

has been publicly reported about informing, but a similar pattern emerges. The former

director of the Pentagon’s Joint Improvised Explosive Device Defeat Organization (JIEDDO)

reported six years into the Afghanistan war that in cases where “[C]oalition forces separate

the enemy from the people... [w]e see tips go way up; we see bomb makers turned in; we see

IED networks dissolve” (Tata, 2007). Reports from the field indicate that “IED finds were

due directly to local national tips” (Cuomo and Gorman, 2010).

Across these conflicts, insurgents are reported to have taken actions to reduce inform-

ing through anonymous channels. In an apparent “effort by the insurgency to tie up the

lines” in Iraq, prank callers attempted to overwhelm call center operators by “berat[ing] and

threaten[ing them]. Women called to offer the operators sex or, they said, just to chat”

(Semple, 2006). Consistent with this story, internal U.S. Government documents reveal that

while calls to tips lines exceeded 5, 000 per day, “[t]hree out of four phone calls [received

were] harrassment or death threats” (U.S. Central Command, U.S. Department of Defense,

2007). Similarly, billboards, which were used to promote tips lines in Iraq, “‘hit a nerve with

the insurgents’ who regularly vandalize[d] [those] promoting the campaign” (Miles, 2004). In

Afghanistan, the Taliban are reported to have “issued decrees ordering all cell phone towers

to be turned off during nightly hours, in an attempt to prevent villagers from calling in tips

to the military forces (Trofimov, 2010) and have attacked and destroyed cell phone towers

for the same purpose (Shachtman, 2008)” (Shapiro and Weidmann, 2015).

To summarize, government forces and their allies in both Afghanistan and Iraq made in-

7

vestments which suggested they placed a great value on information flow from non-combatants

and insurgents in both conflicts behaved as though they thought informing posed a significant

threat.

2.2 Heterogeneous Response to Civilian Casualties

Intuitively state- and insurgent-caused civilian casualties might be predicted to operate in

opposite ways for a number of reasons, and we are agnostic as to which mechanism is

most important. State-caused casualties may adversely affect citizens’ attitudes about the

government, decreasing their willingness to share tips. Those caused by the insurgency might

induce analogous sentiments towards insurgents and therefore increase tips to government.

The desire for revenge, particularly amongst civilians who suffer the loss of family member

or friend, might also affect willingness to inform. As Schumann and Ross (2010) explain,

“individuals experience distress when they have been treated unfairly... [and] [r]evenge may

enable victims to reduce their distress by restoring equity with the transgressor.” Revenge

can also provide the victim with a sense of psychological gratification (Crombag et al., 2003)

in ways that are highly valued (DiGiuseppe and Tafrate, 2007). One way to exact revenge

in an internal conflict is to help the enemy of the party that has hurt you.

In the context of civil wars like the one in Iraq, however, the actions taken to harm those

who have hurt you carry asymmetric risk depending on whether that party is the insurgency

or the government and its allies. Civilians harmed by the state can simply abstain from

informing, enabling insurgents to conduct more attacks. Civilians who suffer at the hands of

insurgents, however, must take a risky action to inform. Although revenge motives in some

settings are found to be relatively insensitive to consideration of the potential risks (O’Connor

and Adams, 2013; Carlsmith et al., 2002), the risks of informing versus abstaining from doing

so were starkly different in Iraq. Not informing in Iraq was the normal thing to do; the vast

8

majority of citizens did not inform, and there was no punishment for abstention.5 Informing,

however, could be quite risky. Although anonymous tips platforms allowed civilians to

discreetly and relatively easily submit information about insurgents, such activities still

carried with them some expected cost – if insurgents were successful in determining that an

individual had informed (by, for instance, checking outgoing call records on his/her phone at

improvised checkpoints), he or she would almost certainly be tortured and killed.6 We would,

therefore, expect a greater level of harm from insurgents would be required to generate the

same reaction.7

Existing evidence is consistent with this expectation. Condra and Shapiro (2012) show

there were asymmetric effects of civilian casualties in Iraq depending on whether the pop-

ulation in a given area was largely pro-government or pro-insurgency and argue that, in

general, “killings by one side [can] have a different impact on subsequent violence than

killings by the other.” Lyall et al. (2013a) provide further evidence of such phenomenon,

finding that: (a) self-reported harm inflicted by the International Security Assistance Force

(ISAF) in Afghanistan is correlated with decreased support for ISAF and increased support

for the Taliban; and (b) while reports of Taliban-inflicted harm have the analogous relation-

ship, it is substantively much smaller and statistically insignificant. As we show, a similar

asymmetry in magnitude existed in Iraq during the study period.

2.3 Expectations

Given this background, we articulate two simple expectations:

5Contrast this with the former East Germany, for example, where informing was pervasive and failure toparticipate was suspicious.

6In survey data used by the authors in separate research, citizens across Baghdad were asked duringthe war whether they had ‘witnessed criminal behavior’ within the past year and informed authorities inresponse. Amongst those citizens who had witnessed such behavior but elected not to inform, the largestnumber of them cited concern over revenge.

7It is also plausible that different segments of the population respond in fundamentally different ways tocounterinsurgent-caused and insurgent-caused casualties. Were this the case, differential effects in reportingmight be based in the different responses of these population segments to both types of casualties.

9

1. Coalition killings of civilians during combat incidents will lead to a decrease in tips

while insurgent killings of civilians in such incidents will lead to a greater flow of tips.

2. The reaction to killings will be asymmetric, tip flow will drop more following a given

level of Coalition-caused harm than it will increase following the same level of insurgent-

caused harm.

Other expectations are reasonable, though not grounded in the literature on informing

in civil war. For example, suppose that citizens who observe civilian casualties are less likely

to leave their homes and as a result tip less often because they literally have fewer possible

tips – remaining indoors reduces opportunities to witness reportable insurgent activity. We

would then see a negative association between changes in Coalition-caused civilian casual-

ties and informing. But, were this dynamic driving the relationship then we would also

expect a negative correlation between insurgent-caused civilian causalities and tips from the

population. As we will see, that was not the case.

3 Data

To test the relationship between civilian casualties and information flow in Iraq we take

advantage of several kinds of administrative data collected by Iraqi and U.S. forces as well as

data on civilian casualties developed by Iraq Body Count, a non-governmental organization

that has been tracking civilian casualties in Iraq since the start of the war in 2003.

3.1 “Tips” Data

Information supplied by local citizens on insurgents and/or their activities was collected and

quantified by various elements of the Coalition and Government of Iraq throughout much of

the war. Tips were collected from Iraqi citizens in a variety of ways.

10

A plurality flowed into the “Iraqi National Tips 1-3-0 Hotline [hereafter, ‘130 hotline’],

an anonymous telephone hotline for reporting terrorist related activity” (U.S. Central Com-

mand, U.S. Department of Defense, 2007). The 130 hotline was established by Coalition

forces in January 2005 and tranferred to Iraq’s Ministry of Interior in November of 2007”

(Multi-National Corps – Iraq, 2008).8 The hotline was modeled after a similar program

established by the British Army during the Troubles in Northern Ireland. In fact, a team of

“police officers from Northern Ireland were drafted in to help to set up the service and train

an Iraqi team to answer the calls 24 hours a day” (Haynes, 2007). The “[t]ips center,” where

incoming calls were received, was initially located in Iraq’s Al Adnon palace, which was then

part of the International (Green) Zone established by Coalition forces following their invasion

(Multi-National Corps – Iraq, 2007). A subsidiary 130 tips center was also established in

Basrah (Multi-National Corps – Iraq, 2007). While the 130 hotline was established to serve

the whole of Iraq, in practice, the large majority (“90%”) of legitimate tips reported through

the hotline were specific to activities within the Baghdad area (Multi-National Corps – Iraq,

2007).

In addition to the 130 hotline, a number of regional channels were established to collect

tips. Between “30 and 60” local and regional tips hotlines also operated during the conflict

(Multi-National Corps – Iraq). Other data was collected by e-mail (Multi-National Corps

– Iraq, 2008) or “initiated by a person” (Multi-National Corps – Iraq, 2007). Collectively,

regional tips outnumbered those received through the 130 hotline U.S. Department of Defense

(2007) (Figure 3). Broadly, non-130 hotline sources of tips received by telephone included:

those called into Provincial Joint Coordination Centers; Iraqi police stations; and “[p]ersonal

8Although the program was transferred, the 130 hotline was advertised as a host-nation program, andoperators were Iraqi citizens including local police officers. The “1-3-0 Hotline Program is advertised asan anonymous, good Samaritan program run by Iraqis. [In contrast,] [t]he Rewards Program is almost bydefinition not Iraqi, not anonymous, and not necessarily appealing to good Samaritans. Tip programs donot want to confuse the message advertised, so tips stays away from the Rewards Program with respect topublic image” (Multi-National Corps – Iraq).

11

[c]ell [p]hones or [c]ommanders (Iraqi Police, Iraqi Army, Coalition Forces)” (Multi-National

Corps – Iraq, 2007). Figure 3 shows examples of the reporting that Multi-National Command

Iraq produced using data on these hotlines.

The 130 hotline was advertised to Iraqi citizens in a variety of ways. In addition to bill-

board advertisements mentioned previously, television commercials, leaflets, business cards,

posters, stickers, and even cigarette lighters were all used to advertise tips channels (Military

Information Support Team – Iraq). Figure 4 shows advertisements run to promote the 130

hotline.

[Insert figure 4 about here.]

Our data on tips to Coalition forces in Iraq are derived from various plots showing

the weekly rate of useful tips (as defined below) by province.9 The plots are in recently

declassified reports prepared by Multi-National Corps – Iraq for internal briefings (Multi-

National Corps – Iraq, 2008).10 Figure 5 shows one such plot. To generate estimates of the

underlying data from which the plots were constructed, each plot was digitized with Plot

Digitizer (Huwaldt), which was then used to extract tip counts for each observation of each

plot. The weekly data cover the period June 1, 2007 through June 27, 2008. Similar graphs

were provided for all but five of Iraq’s provinces. Three of the five missing provinces were

in the Kurdish region, where there was very little combat and where Coalition forces were

likely not engaged in efforts to recruit informants. The Muthanna and Dhi Qar provinces,

both of which are relatively peaceful majority Shia provinces, are also missing. These gaps

are unlikely to impact our conclusions. The 13 provinces covered in the data account for

99.1% of the combat incidents recorded by Coalition forces for the period under study and

99.0% of the civilian casualties recorded in the data described below.

9Iraq has 18 provinces whose 2007 population ranged from 544,000 people in Dahuk in the Kurdish regionsto Baghdad with 7.1M people. The median province had 1M people in 2007.

10The documents were released to the authors for the purposes of this and related studies through aFreedom of Information Act (FOIA) request.

12

[Insert figure 5 about here.]

Critically, the tips data we use do not count the raw number of calls placed to the tips

lines—which would be a biased measure of information flow because of the insurgent efforts

to flood the line with fake tips discussed above. Rather, the data record the number of

tips classified by Coalition and Iraqi operators as valuable. Specifically, tips were classified

into two types: “informational”, were defined as consisting of tips for which “[n]o immediate

response [is] required; useful in developing intelligence and further research” (Multi-National

Corps – Iraq, 2008); and “preemptive”, defined as consisting of tips “requir[ing] an immediate

on-the-ground response” (Multi-National Corps – Iraq, 2008). The classification of tips into

these two subcategories does not appear to have been consistent across time and between

provinces. We therefore aggregate up to a measure of total useful tips for our analysis.11

Civilian Casualties

Data on civilian casualties (CIVCAS) come from Iraq Body Count (IBC), a non-profit orga-

nization that collects data on civilian casualties suffered in the Iraq War from press reports.12

The data are based on media reports of incidents involving civilian casualties between De-

cember 2003 and the present and have been previously used in many studies the conflict in

Iraq. We use the subset of these data covering our period.

Civilian casualties in the IBC data can be divided into four different categories following

Condra and Shapiro (2012): (1) insurgent killings of civilians that occur in the course of

attacking Coalition or Iraqi government targets, which we call insurgent-caused casualties;

(2) Coalition forces killings of civilians in the course of combat incidents or attacks on

insurgents, which we call coalition-caused casualties; (3) sectarian killings defined as those

11It is impossible to know what portion of tips were considered useful. Anecdotally our interviews suggestthe level of fake tipping varied greatly over time and between provinces.

12http://www.iraqbodycount.org.

13

conducted by an organization representing an ethnic group which did not occur in the context

of attacks on Coalition or Iraqi forces; and (4) unknown killings, where a clear perpetrator

could not be identified. This last category captures much of the violence associated with

ethnic cleansing, reprisal killings, and the like, where claims of responsibility were rarely

made and bodies were often simply left abandoned.

Critically, the killings in the first two categories do not represent the intentional targeting

of civilians. Rather, they represent accidents of war. The rates with which these killings

happened certainly reflected the level of care the two sides took to avoid them, which varied

from place-to-place on both sides, but they do not include acts of intimidation killing targeted

at specific individuals. Because of the accidental nature of these events they are unlikely

to include the violence which was a result of insurgent strategic action to forestall tips

to Coalition and Iraqi forces. That kind of violence would be captured in the latter two

categories of civilian casualties.

3.2 Covariates

Incidents of insurgent violence carried out against and recorded by Coalition forces are

taken from “significant activity” (SIGACT) reports by Coalition forces that capture the

location, date, time, and type of attack for attacks targeted against coalition, Iraqi Security

Forces, civilians, Iraqi infrastructure and government organizations.13 Population data are

extracted from World Food Programme, UN (2008) surveys and the Landscan 2008 dataset,

which provides girded population estimates at 30 arc-second resolution (approximately 500m

x 500m) for the entire world.

13For further details see Berman et al. (2011).

14

3.3 Descriptive Statistics

Tip flow was substantial throughout the period under study, with Coalition forces receiving

24 useful tips in the median province/week. The numbers were highly variable, however, with

some weeks exhibiting exceptionally high numbers of tips and others reporting no tips at all.

As we show below, this skewed distribution does not drive our results. The time series were

also extremely idiosyncratic across provinces, both in terms of levels and the shape of the

time series, as figure 6 shows. We plot the number of tips per combat incident to normalize

for the fact that the intensity of insurgent activity varied greatly across provinces. Given

the skewed distribution of tips we show that all results are statistically and substantively

similar if we trim the tips time series at 3 standard deviations from the mean.14

[Insert figures 6 and 7 about here.]

Insurgents were responsible for roughly twice as many civilian casualties in an average

week as Coalition forces, 3.3 vs. 1.7 as table 1 shows. As is clear from the skewed dis-

tribution, and as can be seen in figure 7, the time series of civilian casualties was highly

episodic.15 There are weeks when both sides cause high numbers of civilian casualties, as

well as weeks when only one side did, and these patterns are imperfectly correlated with

overall combat intensity. The bivariate weekly correlation between combat incidents and

Coalition-caused civilian casualties is roughly .46, and .49 with insurgent-caused casualties.

These correlations, however, are heavily influenced by secular trends in the conflict. In dif-

ferences the correlations drop substantially, to .29 for Coalition-caused incidents and to .07

for insurgent-caused ones.

[Insert table 1 about here.]

14The results of the trimmed regressions are insensitive to trimming at +/− 3 s.d. from the medianinstead.

15Because we do have no strong reasons to expect substantial non-linearities in the response to civiliancasualties and because both civilian-casualties time series exhibit a large number of zeros we do not log them.

15

4 Empirical Strategy

To test the prediction that insurgent- (Coalition-)caused civilian casualties increased (de-

creased) information flow, we estimate the relationship between civilian casualties incident

to combat operations and subsequent tip flow. Because telephone and e-mail tips channels

would allow individuals to respond rapidly to observed civilian casualties, we adopt the

province-week as our level of analysis. We expect to find that the number of tips responds

to casualties from the previous week.

As figures 6 and 7 showed there were strong trends in both time series. We therefore

estimate the following first-differences model:

(yj,t−yj,t−1) = α+β1(ccj,t−1−ccj,t−2)+β2(c

ij,t−1−cij,t−2)+β3(c

sj,t−1−csj,t−2)+γt +(εj,t− εj,t−1),

where, yj,t is the number of tips in province j during week t, ckj,t is the count of civilian

causalities in province j during week t attributable to source k ∈ {Coalition (c), insurgents

(i), sectarian (s)}. Week fixed effects, γt, control for linear responses to unobserved char-

acteristics of each particular week (e.g. major terrorist attacks like the March 6 bombing

in Baghdad’s Karrada neighborhood which killed 68, wounded 120, and might have affected

informing throughout the country). Differencing period t and period t − 1 in this model

eliminates permanent unobserved trends in the individual provinces.16

Given that we are analyzing the weekly response of tips to weekly changes in killings,

omitted variable bias from factors such as endogenous shifts in Coalition and insurgent

capabilities (i.e. force levels) could only drive our results if those capabilities varied across

very short time horizons. That is implausible given what we know about the force allocation

16In levels, both time series exhibit non-stationarity. In differences, this is not the case whether we lookprovince-by-province using a Box-Pierce test or whether we use the common panel data unit root tests.Details available from authors.

16

process on both the Coalition and insurgent sides.17 Controlling for lagged changed in

sectarian killings should account for intimidation killings by insurgents designed to prevent

the population from cooperating with Iraqi and Coalition forces.

We cluster standard errors at the province to take into account the potential for serial

correlation in the error term within provinces when assessing statistical significance. Be-

cause our dataset cover only 13 provinces, we assess statistical significance using the wild

cluster bootstrap procedure developed by (Cameron et al., 2008) (hereafter, CGM), which

allows for cluster-robust inference when the number of data groups is small. With only

thirteen provinces we are well above the minimum of six clusters suggested by CGM based

on monte carlo simulations. Because this procedure does not return an analytical p-value

we report 95% confidence intervals throughout. In the appendix we also report standard

heteroskedasticity-robust standard errors clustered at the province.

Substantively, this approach follows Condra et al. (2010) and Condra and Shapiro (2012)

who leverage the randomness inherent in weapons effects to identify the impact of civilian

casualties. Simply put, the consequences of any given combat engagement have a large

random component; civilians sometimes suffer for being at the proverbial wrong place at

the wrong time. We attempt to isolate the causal impact of those idiosyncratic week-to-

week changes in civilian casualties from fighting between Coalition and insurgent forces by

estimating first differenced regressions with a time fixed effect to account for secular trends

in the war.

Identifying off of this randomness allows for causal inference. The idea is that differenc-

ing out the unit specific trends should remove any omitted variables that impact both the

propensity for civilians to suffer in t−1 and the rate of information flow in t (e.g. differences

17On the Coalition side the authors have interviewed many officials responsible for force allocation decisionsin 2006-8 and found no evidence of week-to-week responses to insurgent actions in terms of moving units.On the insurgent side captured documents from some of the major insurgent groups operating during thisperiod do not reflect a great deal of personnel mobility. On the latter point see documents on “AQ and OtherSunni Jihadist Groups in Iraq” here: https://www.ctc.usma.edu/programs-resources/harmony-program.

17

in the progress of the Sunni Awakening across provinces). Because our analysis in this paper

is at the province level, the first level administrative unit, as opposed to the district level,

the lower level unit used in much of the prior work on the Iraq war, some of the randomness

inherent in combat may be averaged out. We therefore employ a rich set of controls to show

that the coefficients change little as we address various potential sources of bias.

Our first-difference estimator will be consistent as long as ∆εj,t = εj,t − εj,t−1 is uncor-

related with changes in casualties from the prior week, ∆ccj,t−1,∆cij,t−1, and ∆csj,t−1. This

assumption will not hold if changes in casualties respond to external events that might be

correlated with changes in tips in the next period. For example, if casualties are a function

of past tips, then tips from the past will be correlated with tips at t and casualties at t− 1.

Mean reversion could also cause a mechanical correlation between tips and casualties: casu-

alties would decrease after a bad event even if tips remain the same, leading to a generally

negative estimate of the correlation.

To address concerns for these types of bias in the first-difference estimator, we implement

a series of controls for feedback from lagged values of yj and ckj . We show that the estimated

coefficients (β1 and β2) change little in magnitude and significance when controlling for

potentially confounding factors from the past. This gives us confidence that casualties are

uncorrelated with the potential sources of feedback. Controlling for past trends in killings

of various kinds, both intentional by insurgents and accidental by both sides, for example,

does little to change the core results which suggests that expectations about future deaths

resulting from informing are unlikely to be driving the results. To address additional concerns

for mean reversion bias, we control for sources of noise in the relationship between civilian

casualties and tips. A priori, mean reversion is more likely to be substantial when causalities

are prone to deviation from their mean. Thus, controlling for factors that plausibly add noise

to the variation in casualties, such as the level of combat in a province at a given time, the

population, and flexible polynomial interactions of these factors, should reduce any spurious

18

results due to mean reversion. As we will see, none of these controls significantly alter the

results.

5 Results

This section discusses our core results and their robustness.

5.1 Main Results

We find direct evidence for our first expectation in section 2.3. As table 2 shows, civilian

casualties perpetrated by Coalition forces in a given week are associated with significant

decreases in the number of tips supplied the week following, and the opposite is true for

insurgent-caused casualties.

[Insert table 2 about here.]

To make it clear where the results are coming from we introduce our key treatment

variables and then controls one column at a time: Column 1 includes only Coalition-caused

killings and Column 2 insurgent-caused ones. Because the two time-series surely covary due

to unobserved factors Column 3 includes both which increases the magnitude and significance

of the anti-insurgent reaction. Column 4 introduces controls for changes in the intensity

of sectarian violence in the previous week, and Column 5 adds in controls for changes in

the intensity of combat in the prior week. To ensure the results are not driven by some

unobserved set of locality specific shocks that span several across weeks, Column 6 introduces

contemporaneous changes in civilian casualties of all three types and combat incidents.18 The

coefficients move around little in terms of substantive or statistical significance in columns

4-6, though the significance of the anti-insurgent reaction falls below the 90% level when we

include controls for contemporaneous violence. In general the results are extremely stable.

18Adding in changes in ”Unknown” casualties does not change any of the results.

19

Our preferred parsimonious specification is presented in column 5 and includes both

Coalition- and insurgent-caused civilian casualties as well as controls for the intensity of

combat and intimidation by insurgents (measure as the previous week’s changes in sectar-

ian violence and combat incidents). In that model, an additional Coalition-caused death

predicts roughly .8 fewer tips the week following.19 Conversely, an additional civilian death

attributable to insurgent action leads to an increase of approximately .5 tips in the follow-

ing week. The difference between these effects is in the expected direction. The reaction

to Coalition-caused deaths is larger but the difference is modest in statistical significance

(t=1.48, p=.16 using standard clustered s.e.). We therefore find modest evidence for our

second expectation in section 2.3. The population’s reaction to Coalition-caused events is

larger than its reaction to insurgent-caused ones.

Interpreting the magnitude of this effect directly from changes in the count of useful tips is

not straightforward as tips themselves vary considerably in quality. Consider, for example,

the various outcomes of tips received and reported by the Defense Department. In some

instances, tips led to important though relatively minor successes, as when, for instance,

“[a]n early morning tip from local residents led members of the 101st Airborne Division’s

1st Brigade Combat Team to an area near Kirkuk, where they found two IEDs” (AFPS,

2006a). In other cases, tips led to very significant discoveries. One such tip, for instance, led

Coalition forces to a “bomb-making factory in western Iraq... [complete with] ‘quite a sizable

selection of chemicals,’ including canisters of chlorine, several 55-gallon barrels of nitric acid

and several bags of fertilizer” (Miles, 2007). Other tips are more difficult to compare. In one

instance, for example, a “tip from an Iraqi motorist led to the capture of a kidnapper and

the freeing of a woman and her four children...” (AFPS, 2006b).

19Approximately 1.3 fewer tips can be expected in the week which has an increase of the average numberof Coalition-caused civilian casualties in a province week (.77×1.7).

20

5.2 Robustness

The results are robust to a series of alternative specifications. First, as discussed in the

previous section, one concern with our approach is that unobserved factors influencing both

trends in past civilian casualties and current tip flow could lead us to misattribute changes

in tips to lagged changes in civilian casualties. Columns 1-3 of table 3 show this concern

is unlikely to be valid. Our core results on insurgent-caused casualties are unchanged when

including four lags of differences in civilian casualties (Column 1), combat incidents (Column

2), and tips (Column 3), while those on coalition-caused casualties become substantively and

statistically stronger.

[Insert table 3 about here.]

A second concern is that our results are simply an artifact of mean reversion or some other

kind of cyclicality in the data. The fact that controlling for lagged changes in tip flow does

not attenuate the results should ease any such concern. Since mean reversion is more likely

to be substantial when causalities are prone to deviation from their mean, we can determine

whether controlling for factors that plausibly add noise to the variation in civilian casualties

attenuates the results. Column (4) does just that, adding what we call size controls which

include the level of combat in a province over the previous eight weeks, the population of

that province, and flexible polynomial interactions of these factors. Adding these controls

again makes the results cleaner. Combining multiple lags of changes plus size controls makes

the result stronger still, as can be seen in Column 5. We, therefore, believe that the effect(s)

of any unobserved feedback processes or mean reversion would be to attenuate the results.

Table 2 represents a conservative presentation of the relationship between civilian casualties

and tip flow.20

20In an alternative approach, we regress changes in tips on an exponentially weighted moving average(EWMA) of changes in civilian casualties. We construct this variable from casualty values over the precedingeight-week period, weighted by the standard decay of e−l−1, where l is the lag order. If our findings were being

21

The results are also robust to addressing various additional potential confounding factors,

as is shown in table 4. The first such concern is that common shocks affecting individual

provinces for fixed periods might induce a correlation between changes in civilian casualties

and changes in tips that is not due to civilians’ response to idiosyncratic civilian casualty

events and is not accounted for by simple differencing. To address this concern, Columns

1-3 control for province or province/period fixed-effects. Accounting for additional time-

invariant differences between locations or time-varying common shocks in this manner makes

little difference in terms of statistical significance, though the coefficient magnitudes do shift

somewhat.

[Insert table 4 about here.]

A third additional source of inaccuracy in the results could be that the model is misspec-

ified because we are treating provinces as equal units when in fact some are much larger.

Under this logic, in larger provinces the relationship between civilian casualties and tips is

drawing on a larger sample of interactions and thus is more precise. If the relationship we

are studying is heterogeneous across differently sized units, the weighted estimates will differ

from the un-weighted estimates.21 Columns 4 and 5 show that results using weighted least

squares based on either population source are consistent with those of the primary specifi-

cation. This should further alleviate any concern over possible model misspecification that

might otherwise result from failure to model heterogeneous population effects (Solon et al.,

2013).

driven by mean reversion, the results should be significantly attenuated by this approach. Instead, resultsunder this specification are consistent with primary model results. When larger decay values are chosen,thus assigning more weight to previous periods, our results attenuate smoothly, as should be expected givenour claim to be identifying off the randomness inherent in short-run conflict dynamics; if our assumptionsabout the data generating process are right than we are effectively averaging in noise by taking a movingaverage.

21Common practice when a quantity is measured with more precision in larger units is to weight bypopulation (see e.g. Greenstone and Hanna, 2014).

22

A fourth concern is that the results may be driven by Shia provinces where there was

a very different dynamic than the insurgency in mixed and Sunni provinces. Column (6)

shows this was not the case. Restricting attention to mixed and Sunni provinces does not

considerably impact the results. The coefficients increase, as might be expected given the

higher rates of insurgent activity in mixed and Sunni areas, and the statistical significance

of the results drops, as it should when the sample size is halved. But the core pattern of

increased tips after insurgent-caused casualties and decreased tips after coalition-caused ones

remains the same.

One might also be worried about outliers driving the results given the highly variable

nature of how many civilian casualties each sides caused in any given week. Columns 7

shows that trimming the civilian casualties time series at +/− 3 standard deviations from

the mean strengthens the results, as does combining trimming with weighting in Column 8.

In many settings, the natural specification to use would be the one that trims to reduce the

influence of outliers and weights to give greater influence to areas with larger populations.

Our core models thus represent conservative estimates.

Outlier provinces are also a potential concern. Table 6 shows the results are not driven by

any particular province. Dropping each province from the sample and rerunning the model

returns results consistent with table 2. Table 7 shows the results are robust to rescaling the

tip numbers for the five Shia-majority provinces which potential made their figures based on

daily averages for the week vs. total flows (the graphs appear to show non-integer numbers).

Overall the results are remarkably stable and robust.

[Insert tables 6 and 7 about here.]

As the appendix tables show the statistical significance of the results is similar when

standard errors for the primary set of regressions and all corresponding alternative variations

introduced above are calculated traditional method of clustering at the province.

23

Finally, we perform a temporal placebo test in which we estimate the conditional correla-

tion between changes in tip flow in period t and changes in Coalition- and insurgent-caused

casualties in period t + 1. If our identifying assumptions hold we would expect past civil-

ian casualties to affect information flow, but we would not expect contemporary values of

information flow to be correlated with future values of accidental civilian casualties in an

obvious way once we control for potential confounders.22 These results, presented in table 8,

show that in all but one of twelve results across six model specifications, leads of civilian

causalities are not significantly correlated with tip flow. The confidence intervals on future

changes in civilian casualties in the main specification, column 6, span zero to a great degree,

while those on lagged civilian casualties, column 7, do not with the same controls.

[Insert table 8 about here.]

6 Conclusion

We have shown that in one very important civil conflict, there is a clear relationship between

harm to civilians and wartime informing. Using newly declassified data on the flow of useful

tips to Iraqi and Coalition forces, we found a robust relationship between civilian casualties

and information flow. Tip flow decreased the week after Iraqi and Coalition forces mistakenly

killed civilians during combat operations and it increase the week after insurgents did so.

In standardized terms the results were meaningful but modest. A one standard deviation

change in Coalition-caused casualties predicted 4.4 fewer tips in the next week (5.7×-.77),

while a one standard deviation change in insurgent-caused casualties led to 3.3 additional

22The potential relationship between information flow today and civilian casualties tomorrow is compli-cated. More tips may lead to more interdictions and thus less attacks, but increases in tips might also(a) enable Coalition forces to operate with more precision and thereby reduce future civilian casualties or(b) make the operational environment harder for insurgents leading them to operate with less precisionthereby increasing casualties. The variation in tips induced by plausibly exogenous variation in past civiliancasualties is not large enough to let us disentangle these effects.

24

tips in the next week (7.0×.47). These changes represent standardized treatment effects

from .055 for insurgent-caused civilian casualties to .074 for Coalition-caused ones. The

magnitude of these effects is modest but meaningful using common metrics for treatment

effects.

A natural question arising from this analysis is: to what set of civil wars are these results

relevant? The Iraq war during the study period was unusual in some respects. The most

obvious difference is in the combat power available to both sides. It is a common trope that

most civil wars involve a dramatic discrepancy in military power, at least in their early stages,

but the scale of the discrepancies in Iraq during this time was unusually large. Because U.S.

and international forces fought with Iraqi government forces, the counterinsurgents forces

writ large were highly mobile (they had ready access to helicopters and heavily armored road

vehicles) and benefited from levels of intelligence support, logistical capacity, and precision

indirect fire power (artillery and air power) that far exceeded what is available to most

states fighting insurgencies. Those capacities enabled them to effectively target any position

in space at nearly any time if they had actionable intelligence. This case may therefore be

an an outlier in terms of the potential for non-combatants to influence the trajectory of the

conflict by sharing information.

Our main contribution is thus to conclusively demonstrate the existence of an informa-

tional channel by which combatant behavior can impact conflict outcomes and to do so

using data that until now has remained classified and thus unavailable for research. The

relationship we observe is not surprising. Many scholars have found indirect evidence of this

phenomenon previously. What is unique in this study is that we directly observe information

flow and are able to benchmark the magnitude of the difference in non-combatant response

to the two sides. Consistent with prior work, we find that government forces pay a higher

price for inflicting the same level of harm, though the difference is statistically modest.

From a policy perspective these results clearly indicate that the U.S. military’s focus

25

in training and doctrine on avoiding harm to civilians is well placed. As Condra et al.

(2010) observe, “in addition to moral and legal concerns, there may be military strategic

value in reducing civilian casualties.” Our results offer quantitative evidence that this is

indeed the case. In contemporary counterinsurgency campaigns marked by the deployment

of anonymous tips platforms, both the insurgency and its state challenger pay a price for

harming civilians.

26

Figures

Figure 1: An Iraqi police officer at work in the “tips” call center established in Baghdad.Source: New York Times. Date (of article): November 5, 2006

27

Figure 2: CIA online tip submission form. The page linking to the weform thanks, in Arabic,those “brave individuals willing to provide information leading to the arrest of terroristsand the leaders of the extremist organizations...” (Author’s translation.) Source: CentralIntelligence Agency. Date (of web access): September 24, 2014

28

Figure 3: Iraq-wide National 130 and regional hotline tip numbers compared for the periodfrom August 2006 to July 2007. Source: U.S. Department of Defense’s Measuring Stabilityand Security in Iraq quarterly report. Date: September 2007.

29

Figure 4: Advertisements for the national 130 tips hotline. Source (for all): U.S. Departmentof Defense. Date (of each): unknown

(a) “Your calls protect Iraq. For the eyes of Iraq... open your eyes.130.” (Author’s translation.)

(b) “For their sake. For the sake of Iraq, open your eyes.”(Author’s translation.)

(c) “For their sake, keep your eyes on Iraq. Call 130.” (Author’s translation.)

30

Figure 5: Information and Pre-Emptive Tips Reported for Diyala Province between June2007 and July 2008. Source: Multi-National Corps – Iraq. Date (of containing report): July21, 2008

31

Figure 6: ‘Useful’ Tip Flow Over Time. Axes are scaled to highlight variation within, ratherthan scale of, information across time. Source: Multi-National Corps – Iraq.

05

1015

20

05

1015

20

02

46

05

10

05

1015

05

1015

20

05

1015

02

46

05

10

010

2030

40

0.5

11.

5

2040

6080

010

2030

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

Anbar Babylon Baghdad Basrah Diyala

Kerbala Missan Najaf Ninewa Qadissiya

Salah al-Din Tameem Wassit

Tip

s pe

r en

emy-

initi

ated

inci

dent

32

Figure 7: Civilian Casualties Over Time. Axes are scaled to highlight variation within,rather than scale of, casualties across time. Source: Iraq Body Count.

010

2030

40

05

1015

20

020

4060

80

010

2030

40

010

2030

40

05

1015

20

02

46

8

0.5

11.

52

010

2030

05

1015

020

4060

05

10

010

2030

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

2007

w26

2007

w40

2008

w1

2008

w14

2008

w27

Anbar Babylon Baghdad Basrah Diyala

Kerbala Missan Najaf Ninewa Qadissiya

Salah al-Din Tameem Wassit

Coalition-caused Insurgent-caused

Civ

ilian

Cas

ualti

es

33

Tables

Table 1: Descriptives

count mean p50 min max sdUseful Tips 780 212 24 0 2930 428∆(Tips) 767 -5.6 -.24 -472 441 59Combat Incidents 780 54.6 18 0 728 91∆(Combat Incidents) 780 -1.2 0 -409 462 32Coalition-Caused Civilian Casualties 780 1.7 0 0 73 5.2∆(Coalition) 780 -.03 0 -64 62 5.7Insurgent-Caused Civilian Casualties 780 3.3 0 0 52 6.2∆(Insurgent) 780 -.01 0 -51 52 7.0Sectarian Violence Casualties 780 15.7 3 0 521 38∆(Insurgent) 780 -.46 0 -504 476 33.6

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Usefultips are calculated from MNC-I briefing slides. Combat incidents are drawn fromMNF-I SIGACT-III database. Civilian casualties are from Iraq Body Count ascoded by the Empirical Studies of Conflict Project (ESOC).

34

Table 2: Impact of Civilian Casualties on Information Flow

(1) (2) (3) (4) (5) (6)Tips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceLagged FD of Coalition-Caused -0.604 -0.723 -0.724 -0.773 -0.748Civilian Casualties (-0.93, -0.28) (-1.03, -0.42) (-1.03, -0.43) (-1.11, -0.44) (-1.15, -0.33)Lagged FD of Insurgent-Caused 0.349 0.468 0.469 0.469 0.429Civilian Casualties (-0.04, 0.74) (0.08, 0.86) (0.08, 0.86) (0.09, 0.85) (-0.02, 0.88)Sectarian casualties No No No Yes Yes YesCombat incidents No No No No Yes NoPeriod t controls No No No No No YesR-Squared 0.075 0.074 0.078 0.078 0.078 0.082Observations 767 767 767 767 767 767Clusters 13 13 13 13 13 13

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculated from MNC-I briefing slides.Combat incidents are drawn from MNF-I SIGACT-III database. Civilian casualties are from Iraq Body Count as coded bythe Empirical Studies of Conflict Project (ESOC). Standard errors are calculated using the CGM wild bootstrap procedure.

35

Table 3: Robustness to Controlling for Past Trends

(1) (2) (3) (4) (5)Tips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceLagged FD of Coalition-Caused -0.718 -0.742 -0.863 -0.766 -1.016Civilian Casualties (-0.97, -0.46) (-1.07, -0.41) (-1.18, -0.51) (-1.11, -0.43) (-1.53, -0.47)Lagged FD of Insurgent-Caused 0.885 0.459 0.449 0.468 0.677Civilian Casualties (0.19, 1.58) (0.07, 0.85) (0.14, 0.76) (0.09, 0.85) (0.14, 1.22)Lags of civilian casualties Yes No No No YesLags of combat incidents No Yes No No YesLags of tips No No Yes No YesSize controls No No No Yes YesR-Squared 0.085 0.080 0.325 0.081 0.333Observations 767 767 715 767 715Clusters 13 13 13 13 13

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculated from MNC-Ibriefing slides. Combat incidents are drawn from MNF-I SIGACT-III database. Civilian casualties are fromIraq Body Count as coded by the Empirical Studies of Conflict Project (ESOC). Regressions with lags addthree lags of changes. Size controls include total violence over previous 8 weeks, sum of squared violenceover last 8 weeks, Landscan population, Landscan population squared, and population interacted with totalviolence. Standard errors are calculated using the CGM wild bootstrap procedure.

36

Table 4: Robustness to Fixed Effects, Weighting, and Trimming

(1) (2) (3) (4) (5) (6) (7) (8)

ProvinceFixed Effects

Province/QuarterFixed Effects

Province/MonthFixed Effects

WFPWeights

LandscanWeights

Sunniand Mixed

Trimmed(No weights)

Trimmed(LS weights)

Lagged FD of Coalition-Caused -0.767 -0.689 -0.819 -0.782 -0.788 -1.116 -1.108 -1.532Civilian Casualties (-1.10, -0.44) (-0.97, -0.41) (-1.14, -0.50) (-0.92, -0.64) (-0.93, -0.65) (-1.67, -0.56) (-1.85, -0.39) (-2.07, -0.99)Lagged FD of Insurgent-Caused 0.469 0.312 0.299 0.689 0.721 0.679 0.604 0.866Civilian Casualties (0.09, 0.85) (0.05, 0.55) (0.03, 0.56) (0.29, 1.09) (0.32, 1.12) (0.06, 1.30) (0.14, 1.08) (0.48, 1.26)R-Squared 0.115 0.110 0.347 0.092 0.092 0.167 0.079 0.097Observations 767 766 766 767 767 354 767 767Clusters 13 13 13 13 13 6 13 13

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculated from MNC-I briefing slides. Combat incidents are drawn fromMNF-I SIGACT-III database. Civilian casualties are from Iraq Body Count as coded by the Empirical Studies of Conflict Project (ESOC). Trimmed regressionstruncate civilian casualty variables at 3 s.d. from the mean. Results are almost identical in substantive and statistical significance if we also trim the tips time-seriesat +/− 3 s.d. from the mean. Standard errors are calculated using the CGM wild bootstrap procedure.

37

Table 5: Eight-Week Exponentially Weighted Average Model

(1) (2) (3) (4) (5) (6)Tips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceWeighted Average of -3.962 -5.599 -5.616 -5.994 -5.664Coalition-Caused Civilian Casualties (-6.31, -1.50) (-7.96, -3.13) (-7.91, -3.19) (-8.67, -3.26) (-9.34, -1.80)Weighted Average of 4.580 5.670 5.683 5.708 6.144Insurgent-Caused Civilian Casualties (-0.19, 9.40) (0.84, 10.64) (0.81, 10.67) (0.88, 10.63) (-0.01, 12.40)Sectarian casualties No No No Yes Yes YesCombat incidents No No No No Yes YesPeriod t controls No No No No No YesR-Squared 0.074 0.075 0.078 0.079 0.079 0.082Observations 767 767 767 767 767 767Clusters 13 13 13 13 13 13

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculated from MNC-I briefing slides.Combat incidents are drawn from MNF-I SIGACT-III database. Civilian casualties are from Iraq Body Count as coded by theEmpirical Studies of Conflict Project (ESOC). Standard errors are calculated using the CGM wild bootstrap procedure.

38

Table 6: Robustness to Dropping Provinces

(1) (2) (3) (4) (5) (6) (7)Anbar Babylon Baghdad Basrah Diyala Kerbala Missan

Lagged FD of Coalition-Caused -0.790 -0.825 -0.640 -0.813 -0.681 -0.802 -0.796Civilian Casualties (-1.11, -0.47) (-1.16, -0.49) (-1.21, -0.09) (-1.14, -0.48) (-1.12, -0.21) (-1.14, -0.45) (-1.15, -0.43)Lagged FD of Insurgent-Caused 0.488 0.482 0.311 0.493 0.399 0.484 0.483Civilian Casualties (0.10, 0.89) (0.09, 0.88) (-0.03, 0.63) (0.10, 0.91) (-0.05, 0.82) (0.10, 0.87) (0.10, 0.87)R-Squared 0.085 0.085 0.082 0.085 0.085 0.085 0.084Observations 708 708 708 708 708 708 708Clusters 12 12 12 12 12 12 12

(8) (9) (10) (11) (12) (13)Najaf Ninewa Qadissiya Salah al-Din Tameem Wassit

Lagged FD of Coalition-Caused -0.784 -0.822 -0.806 -0.760 -0.629 -0.798Civilian Casualties (-1.13, -0.42) (-1.15, -0.50) (-1.15, -0.45) (-1.08, -0.43) (-0.86, -0.39) (-1.12, -0.46)Lagged FD of Insurgent-Caused 0.481 0.439 0.480 0.711*** 0.402 0.483Civilian Casualties (0.10, 0.87) (0.06, 0.80) (0.09, 0.86) (0.32, 1.07) (0.11, 0.70) (0.10, 0.88)R-Squared 0.085 0.083 0.084 0.085 0.097 0.085Observations 708 708 708 708 708 708Clusters 12 12 12 12 12 12

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculated from MNC-I briefing slides.Combat incidents are drawn from MNF-I SIGACT-III database. Civilian casualties are from Iraq Body Count as coded bythe Empirical Studies of Conflict Project (ESOC). Standard errors are calculated using the CGM wild bootstrap procedure.

39

Table 7: Robustness to Rescaling Tip Flow For Provinces that Potentially Reported Averages

(1) (2) (3) (4) (5) (6)Tips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceLagged FD of Coalition-Caused -0.507 -0.631 -0.632 -0.676 -0.661Civilian Casualties (-0.98, -0.03) (-1.06, -0.19) (-1.06, -0.19) (-1.13, -0.20) (-1.14, -0.14)Lagged FD of Insurgent-Caused 0.386 0.489 0.490 0.490 0.374Civilian Casualties (0.01, 0.77) (0.10, 0.89) (0.10, 0.90) (0.11, 0.89) (-0.08, 0.84)Sectarian casualties No No No Yes Yes YesCombat incidents No No No No Yes NoPeriod t controls No No No No No YesR-Squared 0.079 0.079 0.082 0.082 0.082 0.087Observations 767 767 767 767 767 767Clusters 13 13 13 13 13 13

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculated from MNC-I briefingslides. For this table only tips were rescaled for Babil, Basrah, Maysan, Karbala, Najaf, and Wasit, the Shia-majorityprovinces which may have been reporting daily averages instead of weekly totals. Combat incidents are drawn from MNF-ISIGACT-III database. Civilian casualties are from Iraq Body Count as coded by the Empirical Studies of Conflict Project(ESOC). Standard errors are calculated using the CGM wild bootstrap procedure.

40

Table 8: Temporal Placebo Test

(1) (2) (3) (4) (5) (6) (7)Tips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceLead FD of Coalition-Caused -0.448 -0.644 -0.341 -0.302 -0.017 -0.109Civilian Casualties (-1.06, 0.16) (-1.10, -0.20) (-0.90, 0.22) (-0.99, 0.40) (-0.49, 0.47) (-0.50, 0.29)Lead FD of Insurgent-Caused 0.086 0.196 0.079 -0.127 -0.223 -0.038Civilian Casualties (-0.31, 0.48) (-0.17, 0.55) (-0.14, 0.30) (-0.61, 0.36) (-0.45, 0.01) (-0.40, 0.33)Lagged FD of Coalition-Caused -0.455Civilian Casualties (-0.73, -0.17)Lagged FD of Insurgent-Caused 0.453Civilian Casualties (0.03, 0.87)Lags of civilian casualties No Yes No No No Yes YesLags of combat incidents No Yes No No No Yes YesLags of tips No No Yes No Yes Yes YesContemporaneous controls No No No Yes Yes Yes YesR-Squared 0.074 0.089 0.318 0.078 0.326 0.335 0.335Observations 767 767 715 767 715 715 715Clusters 13 13 13 13 13 13 13

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculated from MNC-I briefing slides. Columns1 - 6 put lead changes in civilian casualties on the RHS, column 7 puts shows the lagged changes for comparison. Combat incidentsare drawn from MNF-I SIGACT-III database. Civilian casualties from Iraq Body Count as coded by the Empirical Studies of ConflictProject (ESOC). Standard errors are calculated using the CGM wild bootstrap procedure.

41

References

AFPS. Marines unearth weapons caches; tipsters lead troops to more caches. Armed ForcesPress Service, 2006a.

AFPS. Accident claims marine; troops nab insurgents, free kidnap victims. Armed ForcesPress Service, 2006b.

Mohammed al Qaisi. Iraqi army retakes control of towns near fallujah. Al Shorfa, 2014.

Al Shorfa. Iraq opens new counter-terrorism hotline. Al Shorfa, 2013.

Al Shorfa. Anbar opens ‘al-qaeda hotline’. Al Sharif, 2014.

Charlotte Alfred. Inside the advocacy group that keeps track of syria’s war casualties.Huffington Post, 2014.

Eli Berman, Jacob N Shapiro, and Joseph H Felter. Can hearts and minds be bought? theeconomics of counterinsurgency in iraq. Journal of Political Economy, 119(4):766–819,2011.

A Colin Cameron, Jonah B Gelbach, and Douglas L Miller. Bootstrap-based improvementsfor inference with clustered errors. The Review of Economics and Statistics, 90(3):414–427,2008.

Kevin M Carlsmith, John M Darley, and Paul H Robinson. Why do we punish?: Deterrenceand just deserts as motives for punishment. Journal of personality and social psychology,83(2):284, 2002.

Daniel A Castro. Do psychological operations benefit from the use of host nation media?2007.

Luke N Condra and Jacob N Shapiro. Who takes the blame? the strategic effects of collateraldamage. American Journal of Political Science, 56(1):167–187, 2012.

Luke N Condra, Joseph H Felter, Radha K Iyengar, and Jacob N Shapiro. The effect ofcivilian casualties in afghanistan and iraq. Technical report, National Bureau of EconomicResearch, 2010.

Costs of War Project. Civilians killed and wounded, 2014.

Hans Crombag, Eric Rassin, and Robert Horselenberg. On vengeance. Psychology, Crimeand Law, 9(4):333–344, 2003.

Scott A Cuomo and Brandon Gorman. Csis’s afghanistan ied metrics report does not tellthe whole story. Small Wars Journal, 2010.

42

Raymond DiGiuseppe and RC Tafrate. Understanding anger disorders. New York, NY:Oxford, 2007.

David Galula. Counterinsurgency warfare: theory and practice. Greenwood PublishingGroup, 2006.

Jim Garamone. Trends positive in baghdad, joint staff general says. Armed Forces PressService, 2007.

Michael Greenstone and Rema Hanna. Environmental regulations, air and water pollution,and infant mortality in india. American Economic Review, 104(10):3038–72, 2014.

Deborah Haynes. The Times, 2007.

Florian M Hollenbach and Jan H Pierskalla. a re-assessment of reporting bias in event-basedviolence data with respect to cell phone coverage, 2015.

JA Huwaldt. Plot digitizer. 2009.

Iraq Body Count. Iraq body count, 2008.

Irish Republican Army. Handbook for Volunteers of the Irish Republican Army: Notes onGuerrilla Warfare. Paladin Press, 1985.

Kay Johnson. Civilian deaths in afghanistan war reach new high in 2014: U.n. Reuters,2014.

Stathis Kalyvas. The Logic of Violence in Civil War. New York: Cambridge UniversityPress, 2006.

Stathis N Kalyvas, Peter Lange, Robert H Bates, Ellen Comisso, Peter Hall, Joel Migdal,and Helen Milner. The logic of violence in civil war. 2006.

Jason Lyall, Graeme Blair, and Kosuke Imai. Explaining support for combatants duringwartime: A survey experiment in afghanistan. American Political Science Review, 107(04):679–705, 2013a.

Jason Lyall, Kosuke Imai, and Yuki Shiraito. Coethnic bias and wartime informing. InAPSA 2013 Annual Meeting Paper, 2013b.

Donna Miles. Hotline succeeding in foiling iraqi insurgents. Armed Forces Press Service,2004.

Donna Miles. Iraqi tips lead to bomb factory discovery. Armed Forces Press Service, 2007.

Military Information Support Team – Iraq. MIST-Iraq Briefing. U.S. Central Command,U.S. Department of Defense.

43

Multi-National Corps – Iraq. Point Paper. U.S. Central Command, U.S. Department ofDefense.

Multi-National Corps – Iraq. Measuring Tips. U.S. Central Command, U.S. Department ofDefense, 2007.

Multi-National Corps – Iraq. MNC-I Tips Report. U.S. Central Command, U.S. Departmentof Defense, 2008.

John A Nagl. Learning to eat soup with a knife: counterinsurgency lessons from Malaya andVietnam. University of Chicago Press, 2009.

Kieran O’Connor and Gabrielle S Adams. Affective antecedents of revenge. Behavioral andBrain Sciences, 36(01):29–30, 2013.

Megan Price, Jeff Klingner, Anas Qtiesh, and Patrick Ball. Updated statistical analysis ofdocumentation of killings in the syrian arab republic. Human Rights Data Analysis Group,Geneva, 2014.

Karina Schumann and Michael Ross. The benefits, costs, and paradox of revenge. Socialand Personality Psychology Compass, 4(12):1193–1205, 2010.

Kirk Semple. U.s. backs hot line in iraq to solicit tips about trouble. New York Times, 2006.

Noah Shachtman. Taliban threatens cell towers. Wired, 2008.

Jacob N Shapiro and Nils Weidmann. Is the phone mightier than the sword? cell phonesand insurgent violence in iraq. International Organization, 2015.

Gary Solon, Steven J Haider, and Jeffrey Wooldridge. What are we weighting for? Technicalreport, National Bureau of Economic Research, 2013.

Anthony Tata. Ieds: Combating roadside bombs. Washington Post, 2007.

Yaroslav Trofimov. Cell carriers bow to taliban threat. Wall Street Journal, 2010.

Mao Tse-Tung. On guerrilla warfare. University of Illinois Press, 1961.

U.S. Central Command, U.S. Department of Defense. Point Paper. U.S. Central Command,U.S. Department of Defense, 2007.

U.S. Department of Defense. Measuring stability and security in iraq – september 2007.Washington, DC: Department of Defense, 2007.

Nils B Weidmann. A closer look at reporting bias in conflict event data. American Journalof Political Science, 2015.

World Food Programme, UN. Comprehensive food security and vulnerability analysis. 2008.

44

A Appendix Tables

Table A.1: Impact of Civilian Casualties on Information Flow

(1) (2) (3) (4) (5) (6)Lagged FD of Coalition-Caused -0.604*** -0.723*** -0.724*** -0.773*** -0.748**Civilian Casualties (0.178) (0.176) (0.175) (0.194) (0.256)Lagged FD of Insurgent-Caused 0.349 0.468* 0.469* 0.469* 0.429Civilian Casualties (0.228) (0.223) (0.223) (0.220) (0.299)R-Squared 0.075 0.074 0.078 0.078 0.078 0.082Observations 767 767 767 767 767 767Clusters 13 13 13 13 13 13Sectarian casualties No No No Yes Yes YesNumber of combat incidents No No No No Yes YesPeriod t controls No No No No No Yes

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculatedfrom MNC-I briefing slides. Combat incidents are drawn from MNF-I SIGACT-III database. Civiliancasualties are from Iraq Body Count as coded by the Empirical Studies of Conflict Project (ESOC).Robust standard errors, clustered at the province level are in parentheses. Significance is shown as∗ ∗ ∗p < .01; ∗ ∗ p < .05; ∗p < .10.

45

Table A.2: Robustness to Controlling for Past Trends

(1) (2) (3) (4) (5)Lagged FD of Coalition-Caused -0.718*** -0.742*** -0.863*** -0.766*** -1.016***Civilian Casualties (0.138) (0.191) (0.192) (0.198) (0.313)Lagged FD of Insurgent-Caused 0.885** 0.459* 0.449** 0.468* 0.677*Civilian Casualties (0.396) (0.221) (0.186) (0.219) (0.361)R-Squared 0.085 0.080 0.325 0.081 0.333Observations 767 767 715 767 715Clusters 13 13 13 13 13Lags of civilian casualties Yes No No No YesLags of combat incidents No Yes No No YesLags of tips No No Yes No YesSize controls No No No Yes Yes

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are cal-culated from MNC-I briefing slides. Combat incidents are drawn from MNF-I SIGACT-IIIdatabase. Civilian casualties are from Iraq Body Count as coded by the Empirical Studiesof Conflict Project (ESOC). Regressions with lags add three lags of changes. Size controlsinclude total violence over previous 8 weeks, sum of squared violence over last 8 weeks,Landscan population, Landscan population squared, and population interacted with totalviolence. Robust standard errors, clustered at the province level are in parentheses. Signifi-cance is shown as ∗ ∗ ∗p < .01; ∗ ∗ p < .05; ∗p < .10.

46

Table A.3: Robustness to Fixed Effects, Weighting, and Trimming

(1) (2) (3) (4) (5) (6) (7) (8)Province FE Prov/Quarter FE Prov/Month FE WFP weights Landscan weights Sunni and Mixed Trimmed Trimmed and LS weights

Lagged FD of Coalition-Caused -0.767*** -0.689*** -0.819*** -0.753*** -0.773*** -1.116** -1.108** -1.408***Civilian Casualties (0.195) (0.163) (0.212) (0.125) (0.138) (0.391) (0.408) (0.320Lagged FD of Insurgent-Caused 0.469* 0.312* 0.299* 0.680*** 0.712*** 0.679 0.604** 0.851***Civilian Casualties (0.220) (0.150) (0.165) (0.206) (0.215) (0.408) (0.255) (0.215)R-Squared 0.115 0.110 0.347 0.140 0.138 0.167 0.079 0.142Observations 767 766 766 767 767 354 767 767Clusters 13 13 13 13 13 6 13 13

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculated from MNC-I briefing slides. Combat incidents are drawn from MNF-I SIGACT-IIIdatabase. Civilian casualties are from Iraq Body Count as coded by the Empirical Studies of Conflict Project (ESOC). Trimmed regressions truncate civilian casualty variables at 3s.d. from the mean. Results are almost identical in substantive and statistical significance if we also trim the tips time-series at +/− 3 s.d. from the mean. Robust standard errors,clustered at the province level are in parentheses. Significance is shown as ∗ ∗ ∗p < .01; ∗ ∗ p < .05; ∗p < .10.

47

Table A.4: Eight-Week Exponentially Weighted Average Model

(1) (2) (3) (4) (5) (6)Weighted Average of Coalition-Caused -3.962*** -5.599*** -5.616*** -5.994*** -5.664**

(1.291) (1.349) (1.333) (1.497) (2.330)Weighted Average of Insurgent-Caused 4.580 5.670* 5.683* 5.708* 6.144

(2.760) (2.809) (2.820) (2.796) (4.037)Sectarian casualties No No No Yes Yes YesCombat incidents No No No No Yes YesPeriod t controls No No No No No YesR-Squared 0.074 0.075 0.078 0.079 0.079 0.082Observations 767 767 767 767 767 767Clusters 13 13 13 13 13 13

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculated from MNC-I briefing slides. Combat incidents are drawn from the MNF-I SIGACT-III database. Civilian casualties arefrom Iraq Body Count as coded by the Empirical Studies of Conflict Project (ESOC). Robust standard errors,clustered at the province level are in parentheses. Significance is shown as ∗ ∗ ∗p < .01; ∗ ∗ p < .05; ∗p < .10.

48

Table A.5: Robustness to Dropping Provinces

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13)Province Dropped Anbar Babil Baghdad Basrah Diyala Kerbala Missan Najaf Ninewa Qadissiya S-a-D Tameem WassitLagged FD of Coalition-Caused -0.790*** -0.825*** -0.640* -0.813*** -0.681** -0.802*** -0.796*** -0.784*** -0.822*** -0.806*** -0.760*** -0.629*** -0.798***Civilian Casualties (0.190) (0.203) (0.350) (0.194) (0.275) (0.203) (0.206) (0.203) (0.194) (0.205) (0.182) (0.148) (0.199)Lagged FD of Insurgent-Caused 0.488* 0.482* 0.311 0.493* 0.399 0.484* 0.483* 0.481* 0.439* 0.480* 0.711*** 0.402** 0.483*Civilian Casualties (0.238) (0.228) (0.195) (0.241) (0.256) (0.230) (0.229) (0.230) (0.228) (0.228) (0.217) (0.176) (0.230)R-Squared 0.085 0.085 0.082 0.085 0.085 0.085 0.084 0.085 0.083 0.084 0.085 0.097 0.085Observations 708 708 708 708 708 708 708 708 708 708 708 708 708Clusters 12 12 12 12 12 12 12 12 12 12 12 12 12

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculated from MNC-I briefing slides. Combat incidents are drawn from the MNF-I SIGACT-IIIdatabase. Civilian casualties are from Iraq Body Count as coded by the Empirical Studies of Conflict Project (ESOC). Robust standard errors, clustered at the province level are inparentheses. Significance is shown as ∗ ∗ ∗p < .01; ∗ ∗ p < .05; ∗p < .10.

49

Table A.6: Robustness to Rescaling Tip Flow For Provinces that Potentially Reported Averages

(1) (2) (3) (4) (5) (6)Lagged FD of Coalition-Caused -0.507* -0.631** -0.632** -0.676** -0.661**

(0.250) (0.233) (0.231) (0.252) (0.291)Lagged FD of Insurgent-Caused 0.386 0.489* 0.490* 0.490* 0.374

(0.231) (0.231) (0.232) (0.228) (0.293)Sectarian casualties No No No Yes Yes YesCombat incidents No No No No Yes YesPeriod t controls No No No No No YesR-Squared 0.079 0.079 0.082 0.082 0.082 0.087Observations 767 767 767 767 767 767Clusters 13 13 13 13 13 13

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful tips are calculatedfrom MNC-I briefing slides. Combat incidents are drawn from the MNF-I SIGACT-III database.Civilian casualties are from Iraq Body Count as coded by the Empirical Studies of ConflictProject (ESOC). Robust standard errors, clustered at the province level are in parentheses.Significance is shown as ∗ ∗ ∗p < .01; ∗ ∗ p < .05; ∗p < .10.50

Table A.7: Robustness to Weighting Casualties by District Population

(1) (2) (3) (4) (5) (6)Tips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceTips

DifferenceLagged FD of Coalition-Caused -0.526 -0.597 -0.583 -0.653 -0.659Civilian Casualties (Weighted) (-0.79, -0.26) (-0.83, -0.35) (-0.80, -0.36) (-0.91, -0.39) (-1.00, -0.31)Lagged FD of Insurgent-Caused 0.369 0.477 0.469 0.475 0.441Civilian Casualties (Weighted) (0.01, 0.75) (0.11, 0.86) (0.10, 0.85) (0.10, 0.85) (-0.11, 0.99)Weighted Sectarian casualties No No No Yes Yes YesCombat incidents No No No No Yes NoPeriod t controls No No No No No YesR-Squared 0.074 0.073 0.076 0.076 0.076 0.079Observations 767 767 767 767 767 767Clusters 13 13 13 13 13 13

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Useful are tips calculated from MNC-I briefingslides. Combat incidents are drawn from the MNF-I SIGACT-III database. Civilian casualties are from Iraq Body Count ascoded by the Empirical Studies of Conflict Project (ESOC). Casualties are weighted by the square root of district populationover one thousand, using World Food Programme population estimates. Standard errors are calculated using the CGM wildbootstrap procedure.

51

B Iraq Body Count Data

The IBC data on civilian casualties used in this study were constructed using media reports.Recent scholarship on reporting biases in media focus and possible effects on statistical resultsat the cross-national level may raises concerns about possible non-random measurement errorin the IBC data (Weidmann, 2015; Hollenbach and Pierskalla, 2015). We believe such biasis unlikely to drive our results as all the arguments we could think of for reporting biaswould point in the same direction in terms of tip flow and civilian casualties regardless ofthe party responsible, whereas we find heterogeneous effects of civilian casualties on tip flowdepending on the perpetrator. For example, an area receiving more troops could see bothmore combat and more reporters due to the press embedding process. If the increased troopswere also going out to collect information we would see a spurious positive correlation betweencasualties reported and tips. Alternatively, areas getting increased cell phone coverage mightsee more reporting because it’s easier for news to get out and more tips because it’s safer toinform, again creating a positive correlation between all kinds of civilian casualty reportingand useful tips. Given the stability of our results to including various controls that wouldaccount for some of these common factors we believe it extremely unlikely that this kind ofbias is a problem. Moreover, the reporting underlying the IBC data drew on local press whooperated widely throughout the war, and the intensity of Western reporting seems unlikelyto have varied much week-to-week (which is what would be required to drive spurious resultsgiven our econometric approach).

However, to address general concerns about IBC data quality, we compare those IBC datawith administrative data recently released by the U.S. Department of Defense that identifiesthe target of attacks in Iraq.These newly-released data are based on instances of violenceagainst civilians observed directly by or locally reported to military forces deployed by theUnited States, Iraq’s central government, and their coalition partners during the war. Theseforces were deployed across Iraq, and their reporting was not affected by the availability ofcellular communications technologies or the presence of embedded reporters, both of whichcould lead to mistakes in media reporting. If the variation in the IBC data on non-combatkillings of civilians are broadly consistent with these administrative sources, then concernsabout them having systematic measurement error should be reduced.23

To assess that consistency we regress coalition recorded incidents of attacks on civiliansagainst incidents of attacks on civilians reported in the IBC dataset. If the IBC data havelittle systematic measurement error, then we expect the count of sectarian and unknownincidents in the IBC data to correlate positively with the Defense Department data and toexplain a significant proportion of total model variability. We do not, however, expect suchresults for coalition and insurgent caused incidents that occurred during combat becausecombatants themselves would have been identified as the target in such cases.24

And indeed, as table B.2 that expectation is borne out. As column 7 shows, approx-

23Measurement error orthogonal to our outcome of interest is, of course, an issue of statistical precisionand not one of bias.

24The military’s data were constructed such that only one target is identified per incident, excluding theidentification of any secondary and tertiary targets of any given event.

52

imately 64% of the variation in Defense Department’s data count of incidents targetingcivilians is explained by the model in which sectarian casualties is the only regressor. Therates of coalition- and insurgent-caused casualties are not significantly correlated with ratesof civilian casualties in the administrative data, which is as it should be if those data capturethe equivalent of the IBC sectarian and unknown categories. These results provide greaterconfidence in the validity of the IBC data for measuring the week-to-week within-provincevariation in civilian casualties (which is the variation our analytical strategy is leveraging).

53

Table B.1: Province-Week SIGACTS-IBC Correlations

(1) (2) (3) (4) (5) (6) (7)Incidents of Incidents of Incidents of Incidents of Incidents of Incidents of Incidents of

Civilian Targeting Civilian Targeting Civilian Targeting Civilian Targeting Civilian Targeting Civilian Targeting Civilian TargetingIncidents of Coalition-Caused -0.094 -0.323*Civilian Casualties (0.163) (0.176)Incidents of Insurgent-Caused 0.028 0.021Civilian Casualties (0.191) (0.172)Sectarian Incidents 0.248*** 0.252***

(0.023) (0.025)Unknown Incidents 0.164 0.115

(0.112) (0.114)Sectarian + Unknown 0.234*** 0.424***

(0.018) (0.049)Constant 5.280** 5.230** 2.989*** 5.178** 3.001*** 2.944*** 2.053**

(2.154) (2.198) (0.958) (2.209) (1.102) (1.055) (0.914)Week Fixed Effects Yes Yes Yes Yes Yes Yes NoGovernorate Fixed Effects Yes Yes Yes Yes Yes Yes NoR-Squared 0.534Overall R-Squared 0.726 0.726 0.764 0.728 0.764 0.766Within R-Squared 0.203 0.203 0.314 0.207 0.313 0.320Observations 780 780 780 780 780 780 780Clusters 13 13 13 13 13 13 13

Note: The variable Sectarian+Unknown includes both sectarian and unknown observations in the IBC data. Its significance is evidence that the unknownsin IBC include observations which the SIGACT data classfies as targeting civilians.

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Combat incidents in which civilians were tageted based on MNF-I SIGACT-III database. Civiliancasualties from Iraq Body Count as coded by the Empirical Studies of Conflict Project (ESOC). Robust standard errors, clustered at the province level in parentheses. Significanceshown as ∗ ∗ ∗p < .01; ∗ ∗ p < .05; ∗p < .10.

54

Table B.2: Province-Month SIGACTS-IBC Correlations

(1) (2) (3) (4) (5) (6) (7)Incidents of Incidents of Incidents of Incidents of Incidents of Incidents of Incidents of

Civilian Targeting Civilian Targeting Civilian Targeting Civilian Targeting Civilian Targeting Civilian Targeting Civilian TargetingIncidents of Coalition-Caused -0.146 -1.193**Civilian Casualties (0.352) (0.464)Incidents of Insurgent-Caused 0.389 0.310Civilian Casualties (0.341) (0.362)Sectarian Incidents 0.341*** 0.358***

(0.024) (0.024)Unknown Incidents 0.281 0.149

(0.223) (0.248)Sectarian + Unknown 0.320*** 0.452***

(0.028) (0.060)Constant 21.647*** 18.931*** 13.350*** 21.237*** 13.606*** 12.426*** 8.011*

(5.861) (6.022) (3.450) (5.843) (3.631) (3.328) (3.685)Month Fixed Effects Yes Yes Yes Yes Yes Yes NoGovernorate Fixed Effects Yes Yes Yes Yes Yes Yes NoR-Squared 0.639Overall R-Squared 0.781 0.782 0.852 0.783 0.851 0.860Within R-Squared 0.266 0.271 0.506 0.274 0.500 0.530Observations 182 182 182 182 182 182 182Clusters 13 13 13 13 13 13 13

Note: The variable Sectarian+Unknown includes both sectarian and unknown observations in the IBC data. Its significance is evidence that the unknownsin IBC include observations which the SIGACT data classfies as targeting civilians.

Notes: Unit of analysis is the province-week, 1 June 2007-18 July 2008. Combat incidents in which civilians were tageted based on MNF-I SIGACT-III database. Civiliancasualties from Iraq Body Count as coded by the Empirical Studies of Conflict Project (ESOC). Robust standard errors, clustered at the province level in parentheses. Significanceshown as ∗ ∗ ∗p < .01; ∗ ∗ p < .05; ∗p < .10.

55